Forecast bias is the tendency of a forecast to be consistently too high (over-forecasting) or too low (under-forecasting). It is measured as the sum of forecast minus actual, divided by the sum of actual demand. Unlike random error, bias doesn't cancel out, so it turns directly into excess inventory or lost sales. Planamind measures bias by item, location and forecast horizon, separates the bias in planner overrides from the statistical forecast, and points to the cause so planners can correct it.
What is forecast bias?
Every forecast is wrong by some amount. Forecast error has two parts:
- Random error: sometimes too high, sometimes too low. Over time it roughly cancels out, and safety stock is there to absorb it.
- Bias: a systematic lean in one direction. The forecast is too high month after month, or too low month after month.
Random error is unavoidable. Bias is almost always fixable, because it has a cause: a model that doesn't suit the item, a trend the forecast hasn't caught, or, very often, human adjustments that lean one way.
How to measure forecast bias
The most useful measure is volume-weighted bias as a percentage:
A positive result means over-forecasting; a negative result means under-forecasting. Because it's weighted by volume, a few small items can't distort the total. Two other measures are common:
- Mean error in units: the average of forecast minus actual, useful for a single item over time.
- Tracking signal: cumulative error divided by the mean absolute deviation, used in some tools to trigger an alert when bias persists.
Measure bias at the level where stock is held (item and location) and at the horizon supply acts on. A forecast that's unbiased one month ahead can still be badly biased three months ahead, when orders were placed.
Why accuracy alone hides bias
Accuracy measures such as MAPE and WAPE use absolute error: they count how far off the forecast was, not which way. Two forecasts can have identical accuracy and completely different consequences:
| Actual (4 months) | Forecast | Accuracy | Bias | |
|---|---|---|---|---|
| Forecast A | 100 each month | 110, 90, 110, 90 | 90% | 0% |
| Forecast B | 100 each month | 110, 110, 110, 110 | 90% | +10% |
Forecast A's misses cancel out, and safety stock handles them. Forecast B builds 40 extra units in four months, and keeps building. On an accuracy dashboard the two look the same. That's why bias must be tracked separately.
What forecast bias costs
- Over-forecasting builds excess inventory every cycle, ties up working capital, raises storage costs and, for short-life products, becomes write-off. See how to reduce inventory without putting sales at risk.
- Under-forecasting causes stockouts, lost orders, expediting costs and damaged customer relationships. It also flatters the history: lost sales never appear in the data, so next year's forecast starts too low.
Bias also compounds through the plan. A biased forecast drives biased replenishment orders, biased material requirements and a biased financial plan. See how demand planning connects to inventory and working capital.
Common causes of forecast bias
- Optimistic overrides. Sales targets, hoped-for launches or “just in case” adjustments that consistently lift the forecast.
- Sandbagging. The opposite: forecasts held down so targets look easier to beat.
- The wrong model. A model that misses a trend or seasonality, or treats intermittent demand as smooth.
- Unadjusted one-offs. A promotion, tender or stockout month left in the history, pulling the baseline up or down.
- Constrained history. Sales lost to stockouts make true demand look lower than it was.
How Planamind addresses forecast bias
Planamind's approach is to make bias visible, trace it to its source and give planners the evidence to correct it. It doesn't quietly rewrite the forecast; the planner stays in control.
Bias measured everywhere accuracy is measured
Bias sits alongside FA%, WAPE and MAPE in the accuracy diagnostics and the forecast accuracy report, by item, location and customer, by month and by group. Items are labelled as over-forecasting or under-forecasting, and the Excel accuracy download classifies every item and location. The insight report PDF includes a bias summary.
Bias by forecast horizon
The tracking sheet keeps snapshots of past forecasts and shows bias at each horizon, so you can see whether the forecast was biased when supply actually acted on it, not just one month out.
Human bias separated from model bias
The forecast value added (FVA) report compares every planner override with the statistical baseline for the same item, location and customer. For each one it shows the baseline error, the final error, whether the override improved or degraded accuracy, how consistently it helped, and its bias. It flags collaborators whose overrides rarely help, and highlights overrides that reduced accuracy, lean high and left more than three months of stock, which is where bias is costing cash. Overrides can be put on a watchlist for the next planning cycles.
Root cause, not just a number
Forecast error root-cause analysis identifies why an item missed. When the override error is larger than the baseline error, it names override drift and suggests clearing the override. Where a different model would suit the item better, for example Croston's method for lumpy demand, it says so.
Ana explains it in plain language
On the forecast accuracy view, Ana, the AI planning assistant, leads with accuracy and the direction of bias, and planners can ask Ana which items are most biased and why.
Protecting stock while bias is being fixed
Ana's safety stock suggestions measure variability around the item's own average error, so a persistent lean is not counted as random noise and doesn't inflate the buffer. Fixing the bias itself stays with the forecast, where it belongs.
Planamind doesn't automatically de-bias forecasts, and it doesn't choose models on bias. Model selection uses accuracy, and correction is a planner decision based on the evidence above. We think that is the right design: an automatic correction hides the cause, which usually sits in a process or an override, not in the maths.
A simple monthly routine to reduce bias
- Review bias by item group and location for the horizon that drives supply (for example, three months ahead).
- Sort by the value of the bias, not just its percentage, and start with the biggest items.
- For each, check the FVA report: is the bias in the statistical baseline or in the override?
- If it's the override, discuss it with whoever made it and clear or correct it. If it's the baseline, check the model and clean one-off events from the history.
- Track the same items next month to confirm the bias has gone.
See how it works on your own data. Send us your last 24 months and we'll show you where your forecast is biased, and what it's costing, in a 60-minute readout within 48 hours. Book a live demo.
Frequently asked questions
What is forecast bias?
Forecast bias is a consistent tendency for a forecast to be too high or too low. Positive bias means over-forecasting; negative bias means under-forecasting.
How do you calculate forecast bias?
A common formula is Bias % = Σ(forecast − actual) ÷ Σ actual × 100. A positive value means the forecast was too high on balance; a negative value means it was too low.
What is a good forecast bias?
As close to zero as possible. Many teams treat a few percentage points either way as acceptable noise and investigate anything larger, especially on high-volume or high-value items.
What is the difference between forecast bias and forecast accuracy?
Accuracy measures how far off the forecast was, regardless of direction. Bias measures the direction. A forecast can be 90% accurate and still consistently over-forecast, building stock every cycle.
What is forecast value added (FVA)?
Forecast value added compares the accuracy of the final forecast with the statistical baseline, to show whether each manual adjustment made the forecast better or worse. Planamind's FVA report does this per item, location and customer, including the bias of each override.
Does Planamind correct bias automatically?
No. Planamind measures bias, separates override bias from model bias and identifies the cause, and planners decide what to correct.
